relay data & privacy
relay handles private phone calls, so what happens to your data should be straightforward.
this page explains the most important parts in plain language.
what happens when relay answers a call?
relay uses several systems to perform different parts of the call.
plivo connects and routes the phone call.
livekit helps operate the realtime audio session.
soniox receives caller audio and converts speech into text.
satin receives transcript text and relevant conversation context and determines how relay should respond.
rumik silk converts relay’s response into speech using voice models including mulberry and muga.
not every provider receives every part of a call.
does relay record calls?
yes, relay records calls handled through the assistant.
the recording may contain:
caller speech;
relay’s speech;
the conversation handled through relay;
information voluntarily provided by the caller; and
information required to generate transcripts and summaries.
direct calls that do not pass through relay are not recorded by relay.
when you take over and relay is no longer participating in the conversation, relay’s active ai handling ends.
how long does relay keep raw call audio?
raw call recordings may be stored for up to 30 days.
during this period, recordings may be used for:
transcription;
call history;
debugging;
customer support;
security investigations;
quality evaluation; and
permitted ai-improvement work.
after the applicable retention period, recordings are deleted or irreversibly anonymised from relay-controlled systems, subject to temporary backups, legal requirements, security investigations, and service-provider deletion schedules.
who receives the actual audio?
different systems receive audio for different purposes.
plivo
plivo provides telephone numbers, call routing, telephony, and recording infrastructure.
it may process:
live call audio;
recordings;
phone numbers;
call metadata; and
call-routing information.
livekit
livekit provides realtime voice-session infrastructure.
it may process:
realtime audio;
session information;
participant information; and
technical connection data.
soniox
soniox provides relay’s realtime speech recognition.
it receives caller audio so that spoken language can be converted into text.
soniox may also receive the stored call recording if relay is configured to use soniox for post-call transcription.
does satin receive my call audio?
not for the normal reasoning step.
satin receives the transcript text generated from the caller’s speech, along with relevant conversational context and tool information.
the realtime flow is approximately:
caller audio → soniox → transcript → satin → response text → rumik silk → spoken response
what are muga and mulberry?
muga and mulberry are voice-generation models used through rumik silk.
they convert relay’s generated text into speech.
they are not relay’s main realtime reasoning models.
relay first determines what it should say, then silk produces the spoken response.
relay can use configured voices including:
aisha;
ira;
lucas; and
theo.
what happens after the call?
relay may create:
a durable transcript;
a call summary;
caller context;
the stated reason for calling;
requested actions;
urgency or attention information; and
information needed for your call history.
how is the permanent transcript created?
by default, relay uses a self-hosted whisper model running inside relay’s own infrastructure.
when this configuration is used, the stored recording does not need to be sent to an external transcription company for the post-call transcription step.
relay can also be configured to use soniox async transcription.
if soniox async is enabled, the full stored recording may be sent to soniox for transcription.
who creates the call summary?
relay may use a configured llm to create the final call digest.
depending on the active configuration, this may include:
openai;
google gemini;
groq; or
a self-hosted rumik model.
only the provider configured for that processing path receives the summarisation request.
the summary model receives transcript text rather than the original call recording.
does relay remove my phone number before creating the summary?
relay may exclude separate structured fields that are not needed for summarisation, such as:
relay account identifiers;
caller-number metadata;
recipient-number metadata;
internal account information; and
separate recipient identity fields.
however, this does not mean the transcript itself contains no personal information.
anything spoken during the conversation may naturally appear inside the transcript.
for example, if the caller says:
their name;
phone number;
address;
email;
order number;
account reference;
company name;
appointment details; or
another personal detail,
that information may appear in the transcript sent for summarisation.
we therefore treat call transcripts as potentially containing personal and sensitive information even when separate structured identifiers have been removed.
does relay upload my contacts?
yes.
if you give relay contacts permission, relay uploads relevant contact names and phone numbers.
this helps relay determine whether the incoming caller is already someone you know.
relay does not use your contacts to independently advertise to those people or send them unsolicited messages.
if you remove contact permission, relay stops future contact synchronisation.
previously uploaded contacts may remain until they are deleted or your relay account is deleted.
does relay use calls to improve ai?
in some circumstances, yes.
limited portions of call audio, transcripts, generated responses, summaries, corrections, interaction outcomes, and failed interactions may be used to evaluate, train, fine-tune, test, or improve relay and rumik’s ai and voice systems.
this may help improve:
english understanding;
hindi understanding;
hinglish understanding;
speech recognition;
pronunciation;
voice quality;
caller-intent detection;
turn-taking;
conversational quality;
safety;
reliability;
latency; and
response quality.
authorised rumik personnel or contractors may review limited samples where human evaluation or annotation is necessary.
where reasonably possible, information that is not required for the relevant improvement task is removed, minimised, masked, aggregated, or de-identified.
do third-party ai companies get to train on relay calls?
sending information to a provider so it can perform a service for relay does not automatically give that provider permission to train its own models using relay data.
we do not sell:
call recordings;
transcripts;
uploaded contacts; or
personal information
for ai training.
we do not permit unrelated third parties to use identifiable relay call data to train their own general-purpose or foundation models unless that use is separately disclosed, contractually permitted, lawful, and any required consent has been obtained.
can a human listen to my calls?
authorised rumik personnel or contractors may review limited interactions where reasonably necessary for:
debugging;
customer support;
quality evaluation;
safety evaluation;
annotation;
model evaluation; or
permitted ai improvement.
access should be limited to people who require it for those purposes.
does relay sell my data?
no.
relay does not sell or rent your personal information.
information may be shared with service providers where necessary to operate relay, when you instruct relay to communicate information to a caller, when required by law, or in the other circumstances described in the full privacy policy.
does relay send my calls to mixpanel?
raw call recordings and private transcripts are not intended to be sent to ordinary product-analytics tools such as mixpanel.
mixpanel may instead receive product events such as:
onboarding completed;
screening started;
screening succeeded;
screening failed;
notification opened;
feature used;
setting changed;
app error; or
performance information.
what does firebase receive?
relay may use firebase for services such as:
authentication;
push notifications;
application infrastructure; and
diagnostics.
firebase may process information such as:
account identifiers;
authentication information;
device information;
notification tokens;
technical metadata; and
diagnostic information.
who processes relay data?
rumik
operates relay
account data, contacts, call data, recordings, transcripts, summaries and other service information
plivo
telephony and call routing
live audio, recordings, phone numbers and call metadata
livekit
realtime voice infrastructure
realtime audio and session information
soniox
realtime speech-to-text
caller audio
soniox async
optional post-call transcription
stored call recording when enabled
satin
conversational reasoning
transcripts, conversation context and relevant tool information
rumik silk
text-to-speech
relay’s generated response text
muga / mulberry
voice generation through silk
text and voice-generation configuration
self-hosted whisper
default post-call transcription
stored recording processed within relay infrastructure
configured summary llm
call digest generation
transcript text, which may contain information spoken during the call
firebase
authentication, notifications and application infrastructure
identifiers, authentication data, notification tokens and technical information
mixpanel
product analytics
product usage events and technical metadata
not every provider receives information from every call.
how long are contacts and account data kept?
uploaded contacts may remain associated with your account while relay provides caller-identification functionality.
when you submit a valid account-deletion request, relay aims to delete or irreversibly anonymise associated personal information within 30 days.
this may include:
account information;
uploaded contacts;
stored recordings;
transcripts;
summaries;
call history;
preferences; and
other information associated with your relay account.
limited information may be retained where required for legal compliance, fraud prevention, security, accounting, or legal claims.
can i delete an individual call?
where relay provides a delete-call option, deleting a call will schedule associated stored information for deletion, including where applicable:
call history;
recording;
transcript;
summary; and
related stored call information.
temporary backups and legally required records may remain for a limited period.
can i permanently delete my relay account?
yes.
after a valid account-deletion request, relay aims to delete or irreversibly anonymise associated personal information within 30 days.
some information may be retained where required by law or reasonably necessary for security, fraud prevention, accounting, disputes, or legal claims.
do callers know they are speaking to ai?
relay is designed to identify itself as an ai assistant acting on your behalf.
where appropriate or required, relay may also inform the caller that:
they are speaking with an ai assistant;
the interaction may be recorded;
the conversation may be transcribed; and
information may be processed using ai.
what happens when i take over the call?
when you take over and relay is no longer participating, relay’s active ai handling ends.
information generated during the relay-handled portion may still remain available according to relay’s retention rules, including:
the recording;
transcript;
summary; and
call history.
where can i learn more about the infrastructure?
rumik ai and voice infrastructure:
plivo telephony infrastructure:
livekit realtime infrastructure:
have a privacy question?
contact us at:
for complete legal details, read the relay privacy policy.